simd: mask_ternlog family + contiguous fast path for stride-4 eq masks

[AdaWorldAPI]

What

Two additions to the mask-op family in src/simd_int_ops.rs, both consumed by lance-graph-java's lgj_hop (companion PR on the same branch name there).

1. mask_ternlog::<IMM>(a, b, c, dst) / mask_ternlog_assign::<IMM>(a, b, c)

Any 3-input Boolean function of three masks in one pass — IMM is the VPTERNLOG truth table (named tables in simd::ternlog). AND is the rank-1 member of this family: a consumer spelling a & b & c as two mask_and_assign passes through a scratch write now spends one AND3 pass (one VPTERNLOGQ per 512 bits; polyfill elsewhere via U64x8::ternlog, all arms already present from W1a-#9).

• Tail contract stated precisely: conforming inputs → conforming dst iff IMM is even; every named immediate is; subset-shaped tables inherit mask_andnot's "subset of a" guarantee.
• Tests: all 256 tables × 13 lengths vs an independent bit-serial reference; AND3 == and∘and (non-vacuous — both narrowers must contribute); tail-iff-even with the odd (NOR3) can-it-fire arm; length-mismatch panics for both forms.
• Re-exported via ndarray::simd. W1a family-shape deviation recorded in .claude/blackboard.md (same rationale as mask_andnot, 2026-08-18).

2. Contiguous fast path in eq_u32_strided_to_mask

At stride_bytes == 4 the kernel gathered 16 bounds-checked scalar u32 reads into a temporary array before the U32x16 compare — a copy where a cast belonged. That is exactly the shape of every lane in a facet-major columnar store, so lance-graph-java's minor-10 store was contiguous and still being read as strided. Now one fixed [u8; 64] window per 16 elements (safe try_into, bounds proven up front) → a single vector load. Existing eq_u32_strided_stride4_matches_contiguous_primitive parity test covers it; no unsafe.

Measured

lance-graph-java columnar_hop_bench, 65 536 rows × 32 facets, equivalence asserted before timing, both changes together:

frontier	before	after
classid (3 933)	6 342 µs	1 203 µs
hop2 (6 943)	6 408 µs	1 101 µs
all (65 536)	7 547 µs	1 851 µs

16 MB of lane read per hop: 2.1 GB/s → ~13 GB/s.

Gates

cargo fmt --check clean · cargo clippy --lib --features std -- -D warnings clean · cargo test --lib --features std 2272 passed.

🤖 Generated with Claude Code

https://claude.ai/code/session_016WkNBjHc2e3zuyz9i8qJEv

---

Generated by Claude Code

Summary by CodeRabbit

• New Features

  • Added generic three-input mask logic operations supporting all 8-bit truth tables.
  • Added in-place variants for updating existing masks.
  • Defined consistent behavior for scalar tails and input-length validation.

• Performance

  • Improved contiguous 32-bit equality-to-mask processing with vectorized loads.
  • Reduced hop-selection processing time by combining mask operations into a single pass, with benchmarks showing up to 5.8× faster timings.

• Compatibility

  • Existing behavior for strided inputs remains unchanged.

[coderabbitai]

Caution

Review failed

The pull request is closed.

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Essentials

Run ID: a83dddec-8c01-4b8e-8ad8-f75f678a9c44

📥 Commits

Reviewing files that changed from the base of the PR and between fe509db and 1493a10.

📒 Files selected for processing (3)

• .claude/blackboard.md
• src/simd.rs
• src/simd_int_ops.rs

---

📝 Walkthrough

Walkthrough

The SIMD API adds generic ternary mask operations with scalar-tail handling and validation. The equality-mask implementation adds a contiguous stride-4 vector-load path. Hop selection now uses one three-mask operation.

Changes

SIMD mask operations and equality optimization

Layer / File(s)	Summary
Ternary mask API and validation src/simd_int_ops.rs, src/simd.rs, .claude/blackboard.md	Added mask_ternlog and mask_ternlog_assign with compile-time truth tables, SIMD processing, scalar tails, length checks, public re-exports, tests, and one-pass AND3 hop selection.
Contiguous equality fast path src/simd_int_ops.rs, .claude/blackboard.md	Added bounds-checked 64-byte vector loads for stride_bytes == 4. Other strided inputs retain scalar gathers.

Estimated code review effort: 3 (Moderate) | ~20 minutes

Suggested reviewers: claude

Poem

A rabbit packs three masks in line
With truth-table bits that neatly shine
Tails hop softly, checks stand guard
Fast lanes race across the yard
One AND3 pass completes the run

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

---

_{Comment @coderabbitai help to get the list of available commands.}

[cursor]

Bugbot couldn't run - usage limit reached

Bugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit.

A user or team admin can review and increase usage limits in the Cursor dashboard.

(requestId: serverGenReqId_42508340-991f-44b9-afe4-cfa691036b75)